ForestProtector: An IoT Architecture Integrating Machine Vision and Deep Reinforcement Learning for Efficient Wildfire Monitoring ^††thanks: This work was supported by a grant from Google AI, under the program TensorFlow Faculty Award 2021.

To collect environmental data, we developed three IoT sensor nodes. Each node consists of a BME280 sensor, an MQ2 sensor, a TTGO LoRa32 module, a T1592 sensor, and a power bank (See Table I for sensor specifications). The BME280 sensor measures temperature, humidity, and pressure, while the MQ2 sensor detects smoke particles in the air. T1592 sensors were included to detect precipitation and halt data collection during rain. The TTGO LoRa32 modules continuously transmit sensor data to the central gateway via the LoRaWAN protocol. To protect the IoT nodes in outdoor conditions (e.g., rain and solar radiation), we designed a custom case in SolidWorks and 3D printed it with PETG (Polyethylene Terephthalate Glycol) material. The design allows for mounting the node to the top of a 4-inch diameter PVC tube to be installed at ground-level.

The gateway uses a Lilygo T3S3 module to gather data from the IoT devices via the LoRaWAN protocol. Within the gateway, this data is transmitted in JSON format to the Jetson Nano, which acts as the central processing unit. A RL agent inside the Jetson Nano, trained on the IoT sensor data provided by the end devices, defines the Raspberry Picam v2 camera’s horizontal orientation (ranging from 0°to 360°). Once positioned, the camera, coupled with a computer vision system, detects smoke — an early indicator of wildfire. If smoke is detected, the gateway sends an alert to a cloud server, which then relays the alert to the designated users.

Additionally, all sensor data are transmitted to the cloud, which is received by an Express.js backend and stored in a MongoDB database for analysis and monitoring. A user-friendly React.js frontend deployed on Vercel provides real-time data visualizations and creates push notifications using a socket connection. Moreover, the system architecture, deployed on an AWS EC2 instance, includes an alert system that sends WhatsApp notifications to a designated phone number using the WhatsAppWeb.js API and displays alerts on the dashboard. This system architecture is depicted in Figure 2.

The agent was designed using a Deep Q-Network (DQN) architecture, suitable for environments with discrete action spaces and complex observational patterns. The DQN model’s objective was to learn a policy that maximizes the cumulative reward by selecting the most relevant sector for monitoring based on indicators of potential wildfire risk. The resulting architecture consists of three fully connected layers: an input layer that processes the environment’s state, two hidden layers of 24 neurons with ReLU activations, and an output layer representing the Q-values for each action.

The Q-value function was updated using the Bellman equation, which models the expected cumulative reward for taking action $a$ in state $s$, followed by an optimal policy. The update rule is given by:

where: $Q(s,a)$ is the estimated value of taking action $a$ in state $s$. $\alpha$ is the learning rate, which controls the weight given to new information during the update. $r$ is the reward received after taking action $a$ in state $s$. $\gamma$ is the discount factor, which balances the importance of immediate versus future rewards. $\max_{a^{\prime}}Q(s^{\prime},a^{\prime})$ represents the maximum expected value of future rewards for the next state $s^{\prime}$, considering all possible actions $a^{\prime}$.

Each IoT sensor node provides observations on smoke levels, temperature, barometric pressure, humidity, and water presence, however, only temperature, humidity, and smoke levels were considered for the model. The model computes a weighted sum of these observations to prioritize the sector of the IoT node $i$ with the strongest indicators of potential wildfire presence. This signal strength for each IoT node $i$ is calculated as follows:

where $W_{s}$, $W_{t}$, and $W_{h}$ are the respective weights assigned to smoke, temperature, and humidity. $B_{t}$ and $B_{h}$ are binary values (1 or 0) indicating if thresholds for temperature and humidity have been surpassed. $C_{s}$ represents the smoke level, normalized to a scale from 0 to 1 by dividing by 100.

Based on the calculated signal for each IoT node $i$, the agent selects the action $a^{*}$ that maximizes the expected reward according to Equation 3. This approach allows the agent to focus on IoT nodes with the strongest evidence of wildfire risk, enhancing the model’s responsiveness to critical conditions.

Moving on, the performance of the DQN agent was evaluated through two primary metrics: the moving average of rewards over episodes and the training loss over time. Together, these metrics validate the DQN agent’s ability to effectively learn and optimize its policy within the given environment. The moving average, calculated over a 100-episode window, provides a smooth indicator of the agent’s learning progress. This is formally defined in Equation 4, for a given episode $t$, the moving average of rewards $\overline{R}_{t}$ over a window $w$. In this equation, $R_{k}$ denotes the total reward received in episode $k$.

As for the Q-network, the training loss provides insights into the convergence. It is computed as the mean squared error (MSE) between the predicted Q-value and the target Q-value for each state-action pair in a batch. Given a mini-batch of size $N$, the MSE loss $L$ can be defined as:

To collect the dataset detailed in Table II, we used video editing tools, scraped videos from social media websites, and used generative AI tools. First, we used After Effects to generate smoke effects with varying wind and orientation, which were then fused with videos of forests recorded in La Paz, Bolivia. We also scraped videos from social media using search terms like “wildfire,” “burning forest,” and “wild forest.” Finally, we used the AI tool Haiper [20] to generate wildfire videos from static images.

We used data augmentation to create the final dataset, applying the following transformations: Horizontal Flip, Random Brightness and Contrast, Gaussian Blur, and Gaussian Noise. Each transformation had a 50% probability of being applied sequentially using the Albumentations library. This process resulted in a final dataset of 4,000 videos. For experimentation purposes, all videos were converted to grayscale, cropped to a 1:1 aspect ratio, and resized to 240x240 pixels.

This study developed and evaluated five distinct 3DCNN architectures (S1 to S5) to optimize smoke detection accuracy in video sequences. The first model, S1, used Conv3D layers with MaxPooling3D to capture spatial and temporal features, followed by dense layers for classification. S2 built upon S1 by replacing MaxPooling3D with AveragePooling3D, which smoothed feature extraction and improved accuracy. S3, which maintained the structure of S2 but used grayscale data, achieved the best results during testing across all architectures, reducing computational load in the process. The conversion to grayscale was performed after observing the superior performance of models trained on grayscale data. S4 extended S2 with BatchNormalization to stabilize training and TimeDistributed layers to prepare data for bidirectional LSTM layers, further enhancing temporal feature learning. Finally, S5 replicated S4 but utilized grayscale data. Figure 3 illustrates the architecture of S5, the best-performing model.

We trained the model using the following hyperparameters: learning rate $\alpha=0.001$, discount factor $\gamma=0.75$ (to balance the importance of immediate and future rewards), and rewards $r=5$ for correctly identifying the IoT node with the strongest signal, and $r=-1$ otherwise. To enhance stability during training, an experience replay buffer of size 100,000 was employed to store past transitions. At each step, a mini-batch of 64 transitions was randomly sampled. The epsilon-greedy policy enabled a balance between exploration and exploitation, with $\epsilon$ initialized at 1.0 and decaying at a rate of 0.995 per episode, down to a minimum of 0.01. As illustrated in Figure 4, the model demonstrated significant improvement in its decision-making capability over the episodes. Additionally, the following weights were assigned to the sensor data: $W_{s}=0.6$ (highest priority, as it provides a continuous measure strongly correlated with wildfire risk), $W_{t}=0.3$ (temperature sensor contributes significantly when combined with humidity), and $W_{h}=0.05$ (humidity sensor has the lowest individual weight).

III provides a comparative analysis of the five 3DCNN architectures (S1 to S5) experimented in the project. Our experiments demonstrated the importance of incorporating modules like Average Pooling, bidirectional LSTM layers, and grayscale conversion for enhanced performance. S5 achieved the highest metrics (accuracy = 0.925, F-score = 0.93), demonstrating superior performance. Figure 5 shows the accuracy and loss over epochs, highlighting S5’s convergence and stability. Furthermore, Figure 6 shows classification examples obtained using S5, where it is worth noting that the model misclassified fog as smoke in the second sample. In the inference stage, the video captured by the camera is divided into an 8x4 grid, and each cell is analyzed individually.

The final prototypes of the IoT device and gateway are shown in Figures 7a and 7b, respectively. On November 30th, 2024, we conducted a series of outdoor experiments in La Paz, Bolivia, to test the system’s effectiveness in detecting fire features in real scenarios. As shown in Figure 7c, we installed the gateway in the middle and the IoT nodes at a distance of 3 meters from the gateway. We ignited a fire on a wheelbarrow. To simulate wildfire proximity, we moved the wheelbarrow closer to the IoT sensor nodes, keeping the fire facing outwards. We measured the time between moving the wheelbarrow closer to each node and the web application issuing an alert. These tests are shown in Table IV. The last column of this table shows the average detection time per IoT node induced by the fire on the wheelbarrow.

Our experimental results in open environments demonstrate the effectiveness of the proposed system in detecting fires and generating timely alerts. The DRL agent effectively directed the camera towards the source of the fire, and the 3DCNN model accurately verified the presence of smoke. The DRL agent quickly defined the camera orientation towards the potential fire source in approximately 0.82 seconds. However, the 3DCNN required significantly longer—over 1.7 minutes on average—to process the multiple frame sequences of 224x224 dimensions, highlighting the potential for optimizing the 3DCNN model’s inference time.

While our system showed promising results, its field deployment revealed several areas for improvement in future research. Firstly, using professional sensors for detecting smoke, CO, CO2, and other fire-related gases is crucial. While MQ2 sensors can detect various gases, including H2, LPG, CH4, CO, alcohol, smoke, and propane, their sensing range is limited. Also, they exhibited high sensitivity only for gases in close proximity to the sensor. Secondly, exploring the integration of solar energy could enhance the system’s autonomy. Finally, further validation requires deploying and evaluating the system in diverse real-world environments and at varying distances.